Output Attentions + output hidden states

transformers/modeling_albert.py

@@ -105,6 +105,7 @@ class AlbertAttention(BertSelfAttention):
     def __init__(self, config):
         super(AlbertAttention, self).__init__(config)
 
+        self.output_attentions = config.output_attentions
         self.num_attention_heads = config.num_attention_heads
         self.hidden_size = config.hidden_size 
         self.attention_head_size = config.hidden_size // config.num_attention_heads
@@ -177,7 +178,7 @@ class AlbertAttention(BertSelfAttention):
         projected_context_layer = torch.einsum("bfnd,ndh->bfh", context_layer, w) + b
         projected_context_layer_dropout = self.dropout(projected_context_layer)
         layernormed_context_layer = self.LayerNorm(input_ids + projected_context_layer_dropout)
-        return layernormed_context_layer
+        return (layernormed_context_layer, attention_probs) if self.output_attentions else (layernormed_context_layer,)
 
 
 class AlbertLayer(nn.Module):
@@ -193,25 +194,45 @@ class AlbertLayer(nn.Module):
 
     def forward(self, hidden_states, attention_mask=None, head_mask=None):
         attention_output = self.attention(hidden_states, attention_mask)
-        ffn_output = self.ffn(attention_output)
+        ffn_output = self.ffn(attention_output[0])
         ffn_output = self.activation(ffn_output)
         ffn_output = self.ffn_output(ffn_output)
-        hidden_states = self.full_layer_layer_norm(ffn_output + attention_output)
+        hidden_states = self.full_layer_layer_norm(ffn_output + attention_output[0])
 
-        return hidden_states
+        return (hidden_states,) + attention_output[1:]  # add attentions if we output them
 
 
 class AlbertLayerGroup(nn.Module):
     def __init__(self, config):
         super(AlbertLayerGroup, self).__init__()
         
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
         self.albert_layers = nn.ModuleList([AlbertLayer(config) for _ in range(config.inner_group_num)])
 
     def forward(self, hidden_states, attention_mask=None, head_mask=None):
+        layer_hidden_states = ()
+        layer_attentions = ()
+
         for albert_layer in self.albert_layers:
-            hidden_states = albert_layer(hidden_states, attention_mask, head_mask)
+            if self.output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
+
+            layer_output = albert_layer(hidden_states, attention_mask, head_mask)
+            hidden_states = layer_output[0]
+
+            if self.output_attentions:
+                layer_attentions = layer_attentions + (layer_output[1],)
+
+        if self.output_hidden_states:
+                layer_hidden_states = layer_hidden_states + (hidden_states,)
 
-        return hidden_states
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (layer_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (layer_attentions,)
+        return outputs  # last-layer hidden state, (layer hidden states), (layer attentions)
 
 
 class AlbertTransformer(nn.Module):
@@ -227,11 +248,30 @@ class AlbertTransformer(nn.Module):
     def forward(self, hidden_states, attention_mask=None, head_mask=None):
         hidden_states = self.embedding_hidden_mapping_in(hidden_states)
 
+        all_attentions = ()
+
+        if self.output_hidden_states:
+            all_hidden_states = (hidden_states,)
+
         for layer_idx in range(self.config.num_hidden_layers):
             group_idx = int(layer_idx / self.config.num_hidden_layers * self.config.num_hidden_groups)
-            hidden_states = self.albert_layer_groups[group_idx](hidden_states, attention_mask, head_mask)  
+            layer_group_output = self.albert_layer_groups[group_idx](hidden_states, attention_mask, head_mask)  
+        
+            hidden_states = layer_group_output[0]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + layer_group_output[1]
+
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
         
-        return (hidden_states,)
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
 
 
 ALBERT_START_DOCSTRING = r"""    The ALBERT model was proposed in